Intel-XE Archive on lore.kernel.org
 help / color / mirror / Atom feed
From: Himal prasad Ghimiray <himal.prasad.ghimiray@intel.com>
To: "Matthew Auld" <matthew.auld@intel.com>,
	"Matthew Brost" <matthew.brost@intel.com>,
	"Thomas Hellström" <thomas.hellstrom@linux.intel.com>,
	"Stuart Summers" <stuart.summers@intel.com>
Cc: <intel-xe@lists.freedesktop.org>
Subject: Re: [PATCH 4/5] drm/xe/userptr: Make the repin retry loop non-fatal
Date: Thu, 19 Nov 2026 22:40:18 +0530	[thread overview]
Message-ID: <2be230c8-7ece-403d-b843-9ffe4f952bdb@intel.com> (raw)
In-Reply-To: <20261119-userptr-repin-v1-4-af2e327833a9@intel.com>


On 19-11-2026 16:44, Matthew Auld wrote:
> From: Thomas Hellström <thomas.hellstrom@linux.intel.com>
>
> drm_gpusvm_get_pages() gives up with -EBUSY after a fixed one second
> budget if the CPU mappings keep changing during the repin. That error
> was treated as fatal: the rebind worker killed the VM, exec returned it
> to user-space and the bind ioctl converted it to -ENOMEM.
>
> Instead, time-stamp the first repin attempt of a userptr and, as long as
> the overall timeout (the userptr_repin_timeout_ms sysfs attribute, read
> on each check) hasn't expired, translate -EBUSY into -EAGAIN. Discount
> idle gaps between attempts (for example if the task was stopped via
> SIGTSTP) so resuming does not immediately trip the timeout. Callers drop
> all locks and retry:
>
> - exec and the rebind worker already retry on -EAGAIN.
> - The bind and madvise ioctls restart from the top with vm->lock
>    dropped. Since new vmas are created on each attempt, they track the
>    overall timeout themselves.
> - The pagefault handler repeats the repin.
>
> Only when the overall timeout expires is -EBUSY returned. Drop the
> -EBUSY to -ENOMEM conversion in new_vma() as -EBUSY now means just that.
>
> v2 (Matt):
> - Reset num_ufence at the retry label in xe_vm_bind_ioctl() so user
>    fences are not counted twice across repin retries.
> - Handle -EAGAIN from xe_vm_alloc_madvise_vma() in xe_vm_madvise_ioctl()
>    with vm->lock dropped and repin timeout tracking.
> - Discount idle gaps longer than XE_USERPTR_REPIN_IDLE_MS between
>    attempts and check signal_pending() / return -ERESTARTSYS in ioctl
>    retry loops so stopping and resuming (e.g. Ctrl-Z) works without
>    immediately timing out.
>
> Signed-off-by: Thomas Hellström <thomas.hellstrom@linux.intel.com>
> Signed-off-by: Matthew Auld <matthew.auld@intel.com>
> ---
>   drivers/gpu/drm/xe/xe_exec.c       |  9 +++++--
>   drivers/gpu/drm/xe/xe_pagefault.c  |  4 +++
>   drivers/gpu/drm/xe/xe_userptr.c    | 47 ++++++++++++++++++++++++++++-----
>   drivers/gpu/drm/xe/xe_userptr.h    | 53 ++++++++++++++++++++++++++++++++++++++
>   drivers/gpu/drm/xe/xe_vm.c         | 43 ++++++++++++++++++++++++-------
>   drivers/gpu/drm/xe/xe_vm_madvise.c | 23 ++++++++++++++++-
>   6 files changed, 159 insertions(+), 20 deletions(-)
>
> diff --git a/drivers/gpu/drm/xe/xe_exec.c b/drivers/gpu/drm/xe/xe_exec.c
> index 1949fb103fad..0e6a197a8e85 100644
> --- a/drivers/gpu/drm/xe/xe_exec.c
> +++ b/drivers/gpu/drm/xe/xe_exec.c
> @@ -10,6 +10,7 @@
>   #include <drm/drm_file.h>
>   #include <uapi/drm/xe_drm.h>
>   #include <linux/delay.h>
> +#include <linux/sched/signal.h>
>   
>   #include "xe_device.h"
>   #include "xe_exec_queue.h"
> @@ -380,8 +381,12 @@ int xe_exec_ioctl(struct drm_device *dev, void *data, struct drm_file *file)
>   		xe_validation_ctx_fini(&ctx);
>   err_unlock_list:
>   	up_read(&vm->lock);
> -	if (err == -EAGAIN)
> -		goto retry;
> +	if (err == -EAGAIN) {
> +		cond_resched();
> +		if (!signal_pending(current))
> +			goto retry;
> +		err = -ERESTARTSYS;
> +	}
>   err_hw_exec_mode:
>   	if (mode == EXEC_MODE_DMA_FENCE)
>   		xe_hw_engine_group_put(group);
> diff --git a/drivers/gpu/drm/xe/xe_pagefault.c b/drivers/gpu/drm/xe/xe_pagefault.c
> index 98cce0f12b12..6c4e5e77e1c6 100644
> --- a/drivers/gpu/drm/xe/xe_pagefault.c
> +++ b/drivers/gpu/drm/xe/xe_pagefault.c
> @@ -191,6 +191,10 @@ static int xe_pagefault_handle_vma(struct xe_gt *gt, struct xe_vma *vma,
>   			struct xe_userptr_vma *uvma = to_userptr_vma(vma);
>   
>   			err = xe_vma_userptr_pin_pages(uvma);
> +			if (err == -EAGAIN) {
> +				cond_resched();
> +				continue;
> +			}
>   			if (err) {
>   				xe_pagefault_set_error(pf, XE_PAGEFAULT_ERROR_VMA_USERPTR_PIN);
>   				return err;
> diff --git a/drivers/gpu/drm/xe/xe_userptr.c b/drivers/gpu/drm/xe/xe_userptr.c
> index 9c1dac0fce6f..39806ca9f54a 100644
> --- a/drivers/gpu/drm/xe/xe_userptr.c
> +++ b/drivers/gpu/drm/xe/xe_userptr.c
> @@ -74,6 +74,23 @@ static bool __xe_vma_userptr_lockdep(struct xe_userptr_vma *uvma)
>   #define xe_vma_userptr_lockdep(uvma)
>   #endif
>   
> +/**
> + * xe_vma_userptr_pin_pages() - Pin and map the pages of a userptr vma
> + * @uvma: The userptr vma.
> + *
> + * Pin the pages and set up the dma mappings. If the CPU mappings keep changing
> + * during the repin, drm_gpusvm_get_pages() eventually gives up with -EBUSY.
> + * That is converted to -EAGAIN until the overall repin timeout, measured from
> + * the first attempt since the last successful repin, expires. The caller is
> + * then expected to drop all locks and retry.
> + *
> + * Context: Either vm->lock held in write mode, or in read mode together with
> + * the vma's fault_lock.
> + *
> + * Return: 0 on success, -EAGAIN if the caller should retry with all locks
> + * dropped, -EBUSY if the overall timeout expired, or another negative error
> + * code on failure.
> + */
>   int xe_vma_userptr_pin_pages(struct xe_userptr_vma *uvma)
>   {
>   	struct xe_vma *vma = &uvma->vma;
> @@ -84,19 +101,35 @@ int xe_vma_userptr_pin_pages(struct xe_userptr_vma *uvma)
>   		.device_private_page_owner = xe_svm_private_page_owner(vm, false),
>   		.allow_mixed = true,
>   	};
> +	int err;
>   
>   	xe_vma_userptr_lockdep(uvma);
>   	xe_assert(xe, xe_vma_is_userptr(vma));
>   
> -	if (vma->gpuva.flags & XE_VMA_DESTROYED)
> +	if (vma->gpuva.flags & XE_VMA_DESTROYED) {
> +		uvma->userptr.repin = (struct xe_userptr_repin){};
>   		return 0;
> +	}
>   
> -	return drm_gpusvm_get_pages(&vm->svm.gpusvm, &uvma->userptr.pages, 1,
> -				    uvma->userptr.notifier.mm,
> -				    &uvma->userptr.notifier,
> -				    xe_vma_userptr(vma),
> -				    xe_vma_userptr(vma) + xe_vma_size(vma),
> -				    &ctx);
> +	xe_userptr_repin_begin(&uvma->userptr.repin);
> +
> +	err = drm_gpusvm_get_pages(&vm->svm.gpusvm, &uvma->userptr.pages, 1,
> +				   uvma->userptr.notifier.mm,
> +				   &uvma->userptr.notifier,
> +				   xe_vma_userptr(vma),
> +				   xe_vma_userptr(vma) + xe_vma_size(vma),
> +				   &ctx);
> +	if (err == -EBUSY) {
> +		u32 timeout_ms = READ_ONCE(xe->userptr_repin_timeout_ms);
> +
> +		if (xe_userptr_repin_should_retry(&uvma->userptr.repin,
> +						  timeout_ms))
> +			return -EAGAIN;
> +	}
> +
> +	uvma->userptr.repin = (struct xe_userptr_repin){};
> +
> +	return err;
>   }
>   
>   static struct mmu_interval_notifier_finish *
> diff --git a/drivers/gpu/drm/xe/xe_userptr.h b/drivers/gpu/drm/xe/xe_userptr.h
> index 2a3cd1b5efbb..179f2aca26a3 100644
> --- a/drivers/gpu/drm/xe/xe_userptr.h
> +++ b/drivers/gpu/drm/xe/xe_userptr.h
> @@ -6,6 +6,8 @@
>   #ifndef _XE_USERPTR_H_
>   #define _XE_USERPTR_H_
>   
> +#include <linux/hmm.h>
> +#include <linux/jiffies.h>
>   #include <linux/list.h>
>   #include <linux/mutex.h>
>   #include <linux/notifier.h>
> @@ -20,6 +22,51 @@ struct xe_vm;
>   struct xe_vma;
>   struct xe_userptr_vma;
>   
> +/**
> + * struct xe_userptr_repin - Timeout tracking state for userptr repin retries
> + * @start: Timestamp in jiffies of the first repin attempt, adjusted for idle
> + *	   gaps, or zero if no repin is in progress.
> + * @last: Timestamp in jiffies when the previous attempt failed with -EBUSY,
> + *	  or zero if no retry is pending.
> + */
> +struct xe_userptr_repin {
> +	unsigned long start;
> +	unsigned long last;
> +};
> +
> +/*
> + * If more than XE_USERPTR_REPIN_IDLE_MS elapses between attempts (for example
> + * when the task was stopped via SIGTSTP/Ctrl-Z), discount the idle gap so
> + * waking up does not immediately expire the overall repin timeout.
> + */
> +#define XE_USERPTR_REPIN_IDLE_MS (5 * HMM_RANGE_DEFAULT_TIMEOUT)
> +
> +static inline void xe_userptr_repin_begin(struct xe_userptr_repin *repin)
> +{
> +	unsigned long now = jiffies;
> +
> +	if (!repin->start) {
> +		repin->start = now ?: 1;
> +		return;
> +	}
> +
> +	if (repin->last &&
> +	    time_after(now, repin->last +
> +		       msecs_to_jiffies(XE_USERPTR_REPIN_IDLE_MS)))
> +		repin->start = (repin->start + (now - repin->last)) ?: 1;
> +	repin->last = 0;
> +}
> +
> +static inline bool xe_userptr_repin_should_retry(struct xe_userptr_repin *repin,
> +						 u32 timeout_ms)
> +{
> +	unsigned long now = jiffies;
> +
> +	repin->last = now ?: 1;
> +	return !timeout_ms ||
> +	       time_before(now, repin->start + msecs_to_jiffies(timeout_ms));
> +}
> +
>   /** struct xe_userptr_vm - User pointer VM level state */
>   struct xe_userptr_vm {
>   	/**
> @@ -92,6 +139,12 @@ struct xe_userptr {
>   	 * read: vm->svm.gpusvm.notifier_lock in write mode or vm->resv held.
>   	 */
>   	bool initial_bind;
> +	/**
> +	 * @repin: Timeout state for an in-progress repin retry sequence.
> +	 * Protected by vm->lock in write mode, or in read mode together with
> +	 * the vma's fault_lock.
> +	 */
> +	struct xe_userptr_repin repin;
>   #if IS_ENABLED(CONFIG_DRM_XE_USERPTR_INVAL_INJECT)
>   	u32 divisor;
>   #endif
> diff --git a/drivers/gpu/drm/xe/xe_vm.c b/drivers/gpu/drm/xe/xe_vm.c
> index ff8d355728fc..0c5121361f8e 100644
> --- a/drivers/gpu/drm/xe/xe_vm.c
> +++ b/drivers/gpu/drm/xe/xe_vm.c
> @@ -18,6 +18,7 @@
>   #include <linux/delay.h>
>   #include <linux/kthread.h>
>   #include <linux/mm.h>
> +#include <linux/sched/signal.h>
>   #include <linux/swap.h>
>   
>   #include <generated/xe_wa_oob.h>
> @@ -2782,14 +2783,6 @@ static struct xe_vma *new_vma(struct xe_vm *vm, struct drm_gpuva_op_map *op,
>   
>   		if (xe_vma_is_userptr(vma)) {
>   			err = xe_vma_userptr_pin_pages(to_userptr_vma(vma));
> -			/*
> -			 * -EBUSY has dedicated meaning that a user fence
> -			 * attached to the VMA is busy, in practice
> -			 * xe_vma_userptr_pin_pages can only fail with -EBUSY if
> -			 * we are low on memory so convert this to -ENOMEM.
> -			 */
> -			if (err == -EBUSY)
> -				err = -ENOMEM;
>   		}
>   	}
>   	if (err) {
> @@ -4085,11 +4078,13 @@ int xe_vm_bind_ioctl(struct drm_device *dev, void *data, struct drm_file *file)
>   	struct drm_gpuva_ops **ops = NULL;
>   	struct xe_vm *vm;
>   	struct xe_exec_queue *q = NULL;
> -	u32 num_syncs, num_ufence = 0;
> +	u32 num_syncs, num_ufence;
>   	struct xe_sync_entry *syncs = NULL;
>   	struct drm_xe_vm_bind_op *bind_ops = NULL;
>   	struct xe_vma_ops vops = { .flags = 0, };
> +	struct xe_userptr_repin repin = {};
>   	struct dma_fence *fence;
> +	bool repin_again;
>   	int err;
>   	int i;
>   
> @@ -4122,6 +4117,9 @@ int xe_vm_bind_ioctl(struct drm_device *dev, void *data, struct drm_file *file)
>   	/* Ensure all UNMAPs visible */
>   	xe_svm_flush(vm);
>   
> +retry:
> +	repin_again = false;
> +	num_ufence = 0;
>   	err = down_write_killable(&vm->lock);
>   	if (err)
>   		goto put_exec_queue;
> @@ -4225,6 +4223,7 @@ int xe_vm_bind_ioctl(struct drm_device *dev, void *data, struct drm_file *file)
>   	xe_vma_ops_init(&vops, vm, q, syncs, num_syncs);
>   	if (args->num_binds > 1)
>   		vops.flags |= XE_VMA_OPS_ARRAY_OF_BINDS;
> +	xe_userptr_repin_begin(&repin);
>   	for (i = 0; i < args->num_binds; ++i) {
>   		u64 range = bind_ops[i].range;
>   		u64 addr = bind_ops[i].addr;
> @@ -4240,12 +4239,15 @@ int xe_vm_bind_ioctl(struct drm_device *dev, void *data, struct drm_file *file)
>   		if (IS_ERR(ops[i])) {
>   			err = PTR_ERR(ops[i]);
>   			ops[i] = NULL;
> +			repin_again = err == -EAGAIN;
>   			goto unwind_ops;
>   		}
>   
>   		err = vm_bind_ioctl_ops_parse(vm, ops[i], &vops);
> -		if (err)
> +		if (err) {
> +			repin_again = err == -EAGAIN;
>   			goto unwind_ops;
> +		}
>   
>   #ifdef TEST_VM_OPS_ERROR
>   		if (flags & FORCE_OP_ERROR) {
> @@ -4308,6 +4310,27 @@ int xe_vm_bind_ioctl(struct drm_device *dev, void *data, struct drm_file *file)
>   		up_read(&vm->lock);
>   	else
>   		up_write(&vm->lock);
> +	if (repin_again) {
> +		/*
> +		 * A userptr repin keeps failing due to concurrent CPU mapping
> +		 * changes. New vmas are created on every attempt so the overall
> +		 * timeout is tracked here. Retry with all locks dropped.
> +		 */
> +		u32 timeout_ms = READ_ONCE(vm->xe->userptr_repin_timeout_ms);
> +
> +		if (xe_userptr_repin_should_retry(&repin, timeout_ms)) {
> +			bos = NULL;
> +			ops = NULL;
> +			syncs = NULL;
> +			vops.flags = 0;
> +			cond_resched();
> +			if (!signal_pending(current))
> +				goto retry;
> +			err = -ERESTARTSYS;
> +		} else {
> +			err = -EBUSY;
> +		}
> +	}
>   put_exec_queue:
>   	if (q)
>   		xe_exec_queue_put(q);
> diff --git a/drivers/gpu/drm/xe/xe_vm_madvise.c b/drivers/gpu/drm/xe/xe_vm_madvise.c
> index 31245c6c96d4..4bb544ee168b 100644
> --- a/drivers/gpu/drm/xe/xe_vm_madvise.c
> +++ b/drivers/gpu/drm/xe/xe_vm_madvise.c
> @@ -6,6 +6,7 @@
>   #include "xe_vm_madvise.h"
>   
>   #include <linux/nospec.h>
> +#include <linux/sched/signal.h>
>   #include <drm/xe_drm.h>
>   
>   #include "xe_bo.h"
> @@ -13,6 +14,7 @@
>   #include "xe_pt.h"
>   #include "xe_svm.h"
>   #include "xe_tlb_inval.h"
> +#include "xe_userptr.h"
>   #include "xe_vm.h"
>   
>   struct xe_vmas_in_madvise_range {
> @@ -592,11 +594,13 @@ int xe_vm_madvise_ioctl(struct drm_device *dev, void *data, struct drm_file *fil
>   		.addr = xe_device_uncanonicalize_addr(xe, args->start),
>   		.range = args->range,
>   	};
> +	struct xe_userptr_repin repin = {};
>   	struct xe_madvise_details details;
>   	u16 pat_index, coh_mode;
>   	struct xe_vm *vm;
>   	struct drm_exec exec;
>   	int err, attr_type;
> +	bool repin_again;
>   	bool do_retained;
>   
>   	vm = xe_vm_lookup(xef, args->vm_id);
> @@ -629,6 +633,8 @@ int xe_vm_madvise_ioctl(struct drm_device *dev, void *data, struct drm_file *fil
>   
>   	xe_svm_flush(vm);
>   
> +retry:
> +	repin_again = false;
>   	err = down_write_killable(&vm->lock);
>   	if (err)
>   		goto put_vm;
> @@ -642,9 +648,12 @@ int xe_vm_madvise_ioctl(struct drm_device *dev, void *data, struct drm_file *fil
>   	if (err)
>   		goto unlock_vm;
>   
> +	xe_userptr_repin_begin(&repin);
>   	err = xe_vm_alloc_madvise_vma(vm, madvise_range.addr, args->range);
> -	if (err)
> +	if (err) {
> +		repin_again = err == -EAGAIN;
>   		goto madv_fini;
> +	}
>   
>   	err = get_vmas(vm, &madvise_range);
>   	if (err || !madvise_range.num_vmas)
> @@ -730,6 +739,18 @@ int xe_vm_madvise_ioctl(struct drm_device *dev, void *data, struct drm_file *fil
>   	xe_madvise_details_fini(&details);
>   unlock_vm:
>   	up_write(&vm->lock);
> +	if (repin_again) {
> +		u32 timeout_ms = READ_ONCE(xe->userptr_repin_timeout_ms);
> +
> +		if (xe_userptr_repin_should_retry(&repin, timeout_ms)) {
> +			cond_resched();
> +			if (!signal_pending(current))
> +				goto retry;
> +			err = -ERESTARTSYS;
> +		} else {
> +			err = -EBUSY;
> +		}
> +	}


LGTM

Reviewed-by: Himal Prasad Ghimiray <himal.prasad.ghimiray@intel.com>


>   
>   	/* Write retained value to user after releasing all locks */
>   	if (!err && do_retained)
>

  parent reply	other threads:[~2026-11-19 17:13 UTC|newest]

Thread overview: 20+ messages / expand[flat|nested]  mbox.gz  Atom feed  top
2026-11-19 11:14 [PATCH 0/5] Non-fatal repin retries and NUMA-local rebind worker Matthew Auld
2026-11-19 11:14 ` [PATCH 1/5] drm/xe/vm: restore userptr invalidations on VMA op unwind Matthew Auld
2026-11-19 12:48   ` Thomas Hellström
2026-11-19 16:35   ` Himal prasad Ghimiray
2026-11-19 11:14 ` [PATCH 2/5] drm/xe: Requeue the rebind worker instead of retrying in-line on -EAGAIN Matthew Auld
2026-11-19 12:51   ` Thomas Hellström
2026-11-19 16:29     ` Matthew Auld
2026-11-19 17:26       ` Thomas Hellström
2026-11-19 17:38         ` Matthew Auld
2026-11-19 16:36   ` Himal prasad Ghimiray
2026-11-19 11:14 ` [PATCH 3/5] drm/xe: Add a userptr repin timeout sysfs attribute Matthew Auld
2026-11-19 16:38   ` Himal prasad Ghimiray
2026-11-19 11:14 ` [PATCH 4/5] drm/xe/userptr: Make the repin retry loop non-fatal Matthew Auld
2026-11-19 13:03   ` Thomas Hellström
2026-11-19 17:10   ` Himal prasad Ghimiray [this message]
2026-11-19 11:14 ` [PATCH 5/5] drm/xe: run compute rebind worker on device NUMA node Matthew Auld
2026-11-19 13:11   ` Thomas Hellström
2026-11-19 13:38     ` Matthew Auld
2026-11-19 13:16 ` ✓ CI.KUnit: success for Non-fatal repin retries and NUMA-local rebind worker Patchwork
2026-11-19 13:54 ` ✓ Xe.CI.BAT: " Patchwork

Reply instructions:

You may reply publicly to this message via plain-text email
using any one of the following methods:

* Save the following mbox file, import it into your mail client,
  and reply-to-all from there: mbox

  Avoid top-posting and favor interleaved quoting:
  https://en.wikipedia.org/wiki/Posting_style#Interleaved_style

* Reply using the --to, --cc, and --in-reply-to
  switches of git-send-email(1):

  git send-email \
    --in-reply-to=2be230c8-7ece-403d-b843-9ffe4f952bdb@intel.com \
    --to=himal.prasad.ghimiray@intel.com \
    --cc=intel-xe@lists.freedesktop.org \
    --cc=matthew.auld@intel.com \
    --cc=matthew.brost@intel.com \
    --cc=stuart.summers@intel.com \
    --cc=thomas.hellstrom@linux.intel.com \
    /path/to/YOUR_REPLY

  https://kernel.org/pub/software/scm/git/docs/git-send-email.html

* If your mail client supports setting the In-Reply-To header
  via mailto: links, try the mailto: link
Be sure your reply has a Subject: header at the top and a blank line before the message body.
This is a public inbox, see mirroring instructions
for how to clone and mirror all data and code used for this inbox